Power system operation situation prediction method and system based on GCN-LSTM

Abstract

The application discloses a power system operation situation predication method and system based on GCN-LSTM, and belongs to the technical field of power system operation situation predication, and the method comprises the following steps: obtaining SCADA measurement data and PMU synchronous phasor data and preprocessing, and constructing a node-level multi-modal feature matrix; a weighted graph structure is constructed based on the connection relationship of the power grid topology, and the spatial embedding features of the nodes are extracted through a double-layer graph convolution network; the spatial embedding features of the nodes are taken as input, and the time sequence dynamic features are extracted based on a long short-term memory network; based on a cross-modal attention mechanism, the spatial feature embedding and the time sequence dynamic features are adaptively weighted and fused to obtain the fused spatio-temporal joint features; and the power system operation situation predication result is output based on the spatio-temporal joint features, and error evaluation and abnormality detection are carried out. The application significantly improves the accuracy, robustness and situation predication capability of the power system operation situation predication, and provides strong technical support for the safe and stable operation of the power grid.

Description

Technical Field

[0001] This invention belongs to the field of power system operation status prediction technology, specifically involving a power system operation status prediction method and system based on GCN-LSTM. Background Technology

[0002] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art.

[0003] With the rapid development of new power systems and the energy internet, the scale, operation mode, and data structure of power systems have undergone profound changes, placing higher demands on real-time monitoring, dynamic prediction, and security situation assessment of system operation. Under the traditional architecture, power system operation monitoring mainly relies on low-frequency measurement data collected by SCADA (Supervisory Control and Data Acquisition) and least-squares-based operational situation prediction methods based on power flow equations. However, SCADA data has long sampling periods, low update frequencies, and is greatly affected by noise and measurement errors, making it difficult to reflect rapid dynamic changes and limiting its ability to predict situations under complex operating conditions.

[0004] With the continuous expansion of the deployment scale of synchronous phasor measurement units (PMUs), high-frequency, synchronous, and accurate phasor data provides a new data foundation for power system operation status prediction and early warning. However, SCADA and PMUs differ significantly in sampling frequency, data structure, noise characteristics, and time synchronization mechanisms. Directly fusing the two types of data can lead to feature conflicts, information mismatch, and estimation bias. In addition, communication anomalies, data pollution, and sensor drift may occur during power system operation, causing some data to be distorted or unreliable. Without effective fusion and anomaly suppression mechanisms, the reliability of situation prediction results will be affected, and even system decision-making risks may be triggered.

[0005] The inventors discovered that traditional operational status prediction methods based on linear models or least squares frameworks are unable to simultaneously handle the nonlinear characteristics, topological coupling relationships, and temporal dynamic evolution patterns of the power grid. They also cannot effectively mine the complementary information between SCADA and PMU data. As a result, traditional operational status prediction methods are insufficient in terms of accuracy, real-time performance, and reliability in the context of rapid growth of multi-source measurement data and increasingly complex system operating environments. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a power system operation status prediction method and system based on GCN-LSTM. This method and system can maintain robust output under the conditions of measurement interference, communication packet loss and unreliable data, and can identify abnormal trends and potential risks in advance. This significantly improves the accuracy, robustness and situation prediction capability of power system operation status prediction, and provides strong technical support for the safe and stable operation of the power grid.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the technical solution of the present invention provides a method for predicting the operating status of power systems based on GCN-LSTM, including: Acquire SCADA measurement data and PMU synchronization phasor data and preprocess them to construct a node-level multimodal feature matrix; A weighted graph structure is constructed based on the power grid topology connection relationship, and the spatial embedding features of nodes are extracted through a two-layer graph convolutional network. Using the spatial embedding features of nodes as input, temporal dynamic features are extracted based on a long short-term memory network. Based on the cross-modal attention mechanism, spatial feature embedding and temporal dynamic features are adaptively weighted and fused to obtain fused spatiotemporal joint features; Based on the spatiotemporal joint characteristics, the power system operation status prediction results are output, and error assessment and anomaly detection are performed.

[0008] In at least one embodiment, SCADA measurement data includes bus voltage, active power, and reactive power; PMU synchronization phasor data includes high-frequency phasor amplitude, phase angle, and synchronization sampling signal; Specifically, based on a unified timestamp generation operator, SCADA measurement data with low sampling frequency is mapped to a unified scale; all SCADA measurement data and PMU synchronization phasor data are rearranged and organized into node-level multimodal feature vectors according to bus number, and multimodal measurement sequences are formed in chronological order. Preprocessing includes time synchronization, missing value completion, dual-channel normalization, and multimodal weighted fusion; Specifically, dual-channel normalization includes: using an interval normalization method for SCADA measurement data; and using a z-score normalization method for PMU synchronization phasor data. Multimodal weighted fusion specifically involves: performing multimodal weighted fusion on the node-level features of SCADA and PMU data based on learnable modal weights to generate a unified high-dimensional fused node feature vector; the fused node feature vectors of all nodes form a multimodal node feature matrix; where the fused node feature vector specifically consists of:

[0009] In the formula, Represents the feature vector of the fused node; and Represents the learnable modal parameter weights; This represents the normalized SCADA data; This represents the standardized PMU data.

[0010] In at least one embodiment, a weighted graph structure is constructed based on the power grid topology connection relationship, specifically including: Using the power grid bus as the graph node and the line connection relationship as the graph edge, and integrating line impedance, electrical distance and power flow sensitivity, the power system topology is represented as an undirected graph; the weighted adjacency matrix in the undirected graph uses the reciprocal of the line impedance as the initial weighting method, and obtains the normalized adjacency matrix through symmetric normalization.

[0011] In at least one embodiment, the spatial embedding features of nodes are extracted using a two-layer graph convolutional network, specifically including: The first-layer graph convolutional network aggregates and linearly maps the features of local neighborhood nodes based on the normalized sparse adjacency matrix, thus obtaining the local spatial features of the nodes after the first-layer graph convolutional network. The second-layer graph convolutional network expands the receptive field of nodes based on local spatial features, extracts global spatial features across regions, and obtains the spatial embedding features of nodes through residual enhancement. The propagation rules of the first-layer graph convolutional network are specifically expressed as follows:

[0012] In the formula, Represents the normalized adjacency matrix; Represents the feature matrix of multimodal nodes; This represents the learnable weight matrix of the first layer of the graph convolutional network; Indicates the activation function; This represents the local spatial features obtained by a node after the first layer of the graph convolutional network. The propagation rules of the second-layer graph convolutional network are expressed as follows:

[0013] In the formula, This represents the learnable parameters of the second-layer graph convolutional network; This represents the global spatial features obtained by a node after the second layer of the graph convolutional network.

[0014] In at least one embodiment, extracting temporal dynamic features based on a long short-term memory network specifically includes: The node embedding vectors output by the two-layer graph convolutional network are reorganized into a time series form according to the node index; The spatial embedding matrix is ​​flattened along the node dimension, so that the input of the Long Short-Term Memory Network is formed into a vector at each time step, thus obtaining the input sequence representation; Given the input features at a given time step and the hidden state of the previous step, the Long Short-Term Memory network outputs the hidden state after traversing the complete node sequence. A multi-step prediction structure is adopted to generate a state lag vector for a future set time step in one forward propagation, thereby obtaining the temporal dynamic features after fusing the entire time window.

[0015] In at least one embodiment, adaptive weighted fusion specifically includes: The spatial embedding features are flattened into spatial embedding vectors, and mapped to spatial vectors of the same dimension as the temporal dynamic features through linear transformation, while keeping the dimension of the mapped spatial features the same as that of the temporal dynamic features. Using spatial or temporal features as query vectors and features from another modality as keys and values, the attention weight of temporal features on spatial features is calculated by scaling dot product attention. The spatial modality is then updated using this attention weight to obtain fused features. By leveraging nonlinear activation functions to facilitate the learning of higher-order interactions between modalities, we can obtain fused spatiotemporal joint features.

[0016] In at least one embodiment, the power system operation status prediction result is output based on spatiotemporal joint features, specifically including: projecting the fused spatiotemporal joint features onto the node state space through linear mapping, generating a bus-level prediction matrix from the obtained output vector through dimension reconstruction operation, and obtaining the estimated values ​​of the state variables, expressed as:

[0017] In the formula, Indicates the first The first busbar Estimates of the state variables; This indicates a dimension reconstruction operation; Indicates the number of buses in the system; This represents the state dimension of each bus. The output vector of the linear mapping is represented as:

[0018] In the formula, This represents the output layer weight matrix; It is the bias vector; This indicates the spatiotemporal joint characteristics after fusion.

[0019] Secondly, the technical solution of the present invention also provides a power system operation status prediction system based on GCN-LSTM, including: The data processing module is configured to: acquire SCADA measurement data and PMU synchronization phasor data and perform preprocessing to construct a node-level multimodal feature matrix; The spatial feature extraction module is configured to: construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of the nodes through a two-layer graph convolutional network; The temporal feature prediction module is configured to extract temporal dynamic features based on a long short-term memory network, taking the spatial embedding features of nodes as input. The spatiotemporal fusion module is configured to: adaptively weight and fuse spatial feature embeddings and temporal dynamic features based on a cross-modal attention mechanism to obtain fused spatiotemporal joint features; The output module is configured to output the power system operation status prediction results based on spatiotemporal joint features, and to perform error assessment and anomaly detection.

[0020] Thirdly, the technical solution of the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the power system operation status prediction method based on GCN-LSTM as described in the first aspect.

[0021] Fourthly, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the power system operation status prediction method based on GCN-LSTM as described in the first aspect.

[0022] The beneficial effects of the above-described technical solution of the present invention are as follows: 1) The power system operation status prediction method based on GCN-LSTM of the present invention integrates power grid topology, SCADA low-frequency data and PMU high-frequency phasor data, and constructs a spatiotemporal joint modeling framework based on GCN-LSTM. It realizes high-precision estimation of power system state variables and prediction of operation status. It can maintain robust output under the presence of measurement interference, communication packet loss and unreliable data, and can realize early identification of abnormal trends and potential risks. It significantly improves the accuracy, robustness and situation prediction capability of power system operation status prediction, and provides strong technical support for the safe and stable operation of the power grid.

[0023] 2) This invention utilizes a two-layer graph convolutional network to extract the spatial topological relationships between power grid nodes, overcoming the shortcomings of traditional least squares methods in characterizing nonlinear structures; it deeply models the temporal dynamic features of multi-source measurements through long short-term memory networks, accurately capturing slow changes, transient disturbances, and trend evolution in operating states; in terms of multimodal fusion, it performs time synchronization, scale normalization, and feature alignment on SCADA and PMU data, enabling heterogeneous measurements with different sampling frequencies and noise characteristics to achieve unified expression, and jointly encodes spatial and temporal features through a fusion layer, enhancing information complementarity and stability. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a schematic diagram of the power system operation status prediction method based on GCN-LSTM disclosed in Embodiment 1 of the present invention; Figure 2 This is a structural diagram of the power system operation status prediction method based on GCN-LSTM disclosed in Embodiment 1 of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] As described in the background section, the purpose of this invention is to overcome the shortcomings of the prior art and provide a power system operation status prediction method and system based on GCN-LSTM. This method and system can maintain robust output even in the presence of measurement interference, communication packet loss, and unreliable data, and can identify abnormal trends and potential risks in advance. This significantly improves the accuracy, robustness, and situation prediction capability of power system operation status prediction, and provides strong technical support for the safe and stable operation of the power grid.

[0028] Example 1 In a typical embodiment of the present invention, such as Figure 1 and Figure 2 As shown, this embodiment discloses a method for predicting the operating status of a power system based on GCN-LSTM, specifically including the following steps: S1. Acquire SCADA measurement data and PMU synchronization phasor data and preprocess them to construct a node-level multimodal feature matrix; S2. Construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of nodes through a graph convolutional network; S3. Using the spatial embedding features of nodes as input, extract temporal dynamic features based on the Long Short-Term Memory network; S4. Based on the cross-modal attention mechanism, the spatial feature embedding and temporal dynamic features are adaptively weighted and fused to obtain the fused spatiotemporal joint features; S5. Output the power system operation status prediction results based on spatiotemporal joint characteristics, and perform error assessment and anomaly detection.

[0029] The following section provides a detailed explanation of the above-mentioned power system operation status prediction method based on GCN-LSTM, with reference to specific implementation methods.

[0030] S1. Acquire SCADA measurement data and PMU synchronization phasor data and preprocess them to construct a node-level multimodal feature matrix.

[0031] S11. Acquire SCADA measurement data and PMU synchronization phasor data.

[0032] This step aims to construct the multi-source measurement data input required for subsequent operational status prediction. This data does not come directly from sampling by a single device, but rather is uniformly acquired, parsed, synchronized, structured, and encapsulated into a directly processable feature input format from the real-time database of the power system dispatching layer, the SCADA terminal of the station control layer, and the PMU synchronization phasor device. This data generation process is completed collaboratively by multi-threaded scheduling, a protocol parser, a data stream cache manager, and a node mapping module, ensuring that multi-source measurements are input to the subsequent graph neural network framework with high reliability, high temporal consistency, and a high degree of structure in a practical engineering environment.

[0033] In this step, SCADA measurement data has a wide coverage and long sampling period, making it suitable for characterizing steady-state power grid operation information. It includes low-frequency measurement data such as bus voltage, active power, and reactive power. PMU synchronization phasor data has a high sampling frequency and microsecond-level clock synchronization capability, reflecting rapid dynamic changes in the system. It includes high-frequency phasor amplitude, phase angle, and synchronization sampling signals. To ensure data consistency between different sites and measurement devices, this embodiment uses a unified bus numbering system and equipment channel identifiers for indexing both SCADA and PMU measurements, achieving cross-regional and cross-device data fusion and traceability management.

[0034] The acquired SCADA measurement data and PMU synchronization phasor data are aligned according to the same timestamp to form a multimodal dataset of the power system. Specifically, in the time series dimension, this step involves constructing a unified time axis. The system determines the base time scale based on the high-frequency sampling period of the PMU, and then aligns the SCADA measurement data to this unified scale through time interpolation and window mapping, thereby achieving automatic time synchronization of asynchronous sampling data.

[0035] Specifically, a unified timestamp generation operator is defined to generate SCADA measurement data with lower sampling frequencies. Mapped to a uniform scale Specifically, it is expressed as:

[0036] In the formula, This represents the unified timestamp generation operator, which consists of linear interpolation and a dynamic alignment function based on window smoothing.

[0037] To ensure that the measurement format is consistent with the power grid topology in the spatial dimension, this step constructs a bus number table and a physical quantity mapping table, and rearranges all SCADA measurement data and PMU synchronization phasor data according to bus number and organizes them into node-level multimodal feature vectors. For any bus node... At any moment Multi-source measurement is uniformly packaged as:

[0038] In the formula, This represents the original feature dimension, which is dynamically determined based on the number of input modes, including voltage amplitude, phase angle, current, power flow, and the real and imaginary parts of the synchronization phasor.

[0039] The measurement data from all nodes form the system-level input matrix. The measurements are written to the buffer sequentially in chronological order, ultimately forming a complete multimodal measurement sequence, represented as:

[0040] In the formula, Represents the first on a unified time scale A point in time; Indicates the total number of busbar nodes in the power grid; Indicates busbar At any moment The original multi-source measurement vector.

[0041] To improve the reliability of input data, this step incorporates a real-time measurement quality evaluation mechanism during the data generation stage. By detecting noise levels, cross-period change trends, and neighborhood consistency, it automatically removes or marks abnormal measurements. This enables the measurement data to have higher validity, stability, and engineering usability, thus laying a solid foundation for subsequent preprocessing, graph structure construction, GCN spatial modeling, and LSTM dynamic modeling.

[0042] S12. Preprocess the acquired data.

[0043] This step aims to perform multi-stage preprocessing and structured feature construction on the multimodal measurement data of SCADA and PMU, making it a high-quality node feature matrix that meets the input requirements of graph convolutional networks. It mainly includes four core processes: time synchronization, missing value completion, dual-channel normalization, and multimodal weighted fusion.

[0044] S121. Time synchronization correction.

[0045] Because SCADA and PMU have different sampling frequencies, time synchronization correction is still required for SCADA after adopting a unified time axis. Let the original SCADA measurement be... PMU high-frequency measurement is The SCADA value after time synchronization correction is denoted as:

[0046] In the formula, and This refers to the continuous sampling time of SCADA; Represents the linear interpolation coefficients; This represents the SCADA value obtained through interpolation and aligned with the PMU.

[0047] S122. Missing value completion.

[0048] This step employs a joint interpolation strategy of local time trend and electrical neighborhood to complete missing values ​​in the measurement data, ensuring the integrity and temporal continuity of the input features. Specifically, nodes are recorded. At any moment If a certain measurement value is missing, its complete value is denoted as:

[0049] In the formula, It is a node At any moment The final completion value; It is a local trend prediction value based on a time window; It is a correlation prediction value based on the electrical neighborhood; These are the weighted coefficients of the two types of interpolation, satisfying .

[0050] S123. Dual-channel normalization.

[0051] To ensure comparability of different modal data in terms of numerical scale and distribution, this step employs different normalization strategies for SCADA and PMU measurement data to maintain the physical proportion of SCADA and the statistical stability of PMU.

[0052] Specifically, an interval normalization method is used for SCADA measurement data to map measured quantities such as voltage, current, active power, and reactive power to... An interval, to eliminate dimensional differences, is represented as:

[0053] In the formula, This represents the normalized SCADA data; This indicates the SCADA value after time synchronization correction; and These represent the minimum and maximum values ​​of the SCADA feature on the training set, respectively.

[0054] For PMU synchronization phasor data, the z-score normalization method is used to normalize the mean of instantaneous phasor amplitudes and phase angles to zero and normalize the variance to maintain the stability of the measurement distribution, expressed as:

[0055] In the formula, This represents the standardized PMU data; This represents the high-frequency measurement value of the PMU; Represents the PMU feature mean; This represents the standard deviation of the PMU characteristics.

[0056] S124. Multimodal weighted fusion.

[0057] To adaptively adjust the importance of different modal features, this step performs multimodal weighted fusion of node-level features from SCADA and PMU data based on learnable modal weights, generating a unified high-dimensional fused node feature vector, which can be specifically represented as:

[0058] In the formula, Represents the feature vector of the fused node; and This represents the learnable modal parameter weights, which are automatically optimized during the training process.

[0059] Based on this, the fused node feature vectors of all nodes form a multimodal node feature matrix, which is represented as:

[0060] In the formula, This indicates the total number of busbar nodes in the power grid.

[0061] S2. Construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of the nodes through a graph convolutional network.

[0062] After constructing the multimodal node feature matrix, the power system topology needs to be incorporated into the deep learning framework to reflect the electrical coupling relationships between buses. This step involves establishing a physically parameter-weighted graph structure and extracting spatial embedding features using a Graph Convolutional Network (GCN). This ensures that the final node vectors reflect their structural location, electrical neighborhood, and regional propagation patterns within the power grid.

[0063] S21. Construct a weighted graph structure.

[0064] In this step, the power grid bus is used as the graph node and the line connection relationship is used as the graph edge. At the same time, the power grid physical indicators such as line impedance, electrical distance and power flow sensitivity are integrated to represent the power system topology as an undirected graph:

[0065] In the formula, Represents the set of bus nodes; Indicates the line connection relationship between nodes; This represents a line weighted adjacency matrix, whose elements are... Represents a node With nodes The electrical coupling strength, which uses the reciprocal of the line impedance as the initial weighting method, is expressed as:

[0066] In the formula, For connecting nodes With nodes The line impedance, whose absolute value represents the electrical distance. The reciprocal form is used so that the smaller the impedance (the stronger the electrical connection), the larger the corresponding edge weight, which better conforms to the power flow propagation characteristics and effectively enhances the electrical rationality of spatial feature modeling.

[0067] To avoid feature shifts caused by differences in node degree and to improve the numerical stability of graph convolution operations, this step performs symmetric normalization on the weighted adjacency matrix, resulting in a normalized adjacency matrix, specifically represented as follows:

[0068] In the formula, This represents the identity matrix, used to add self-joins to each node; This is the degree matrix. This normalized adjacency matrix ensures that the feature contributions of different nodes are balanced during convolution operations, avoiding over-reliance on nodes with high connectivity and making training more stable.

[0069] To meet the computational demands of large-scale power grids, this step introduces an adaptive sparsity mechanism. A dense-to-coefficient transformation function further converts the adjacency matrix into a CSR sparse storage structure suitable for graph convolution operations. By learning edge weight thresholds, weak connections are automatically eliminated, thereby reducing computational complexity and improving spatial feature propagation efficiency. The resulting weighted graph structure can simultaneously describe the topological characteristics and regional electrical coupling features of the power grid, providing a high-quality graph structure for subsequent spatial convolution.

[0070] S22. Extract the spatial embedding features of the nodes.

[0071] In this step, a two-layer cascaded graph convolutional network (GCN) structure is used to extract the spatial coupling relationship of the power grid at multiple scales to obtain spatial features at different levels.

[0072] The first layer of GCN focuses on capturing the electrical linkage characteristics within the local topological neighborhood. It aggregates and linearly maps the features of local neighborhood nodes based on a normalized sparse adjacency matrix, thereby encoding the local spatial coupling relationships of the bus. Its propagation rules are specifically expressed as follows:

[0073] In the formula, Represents the normalized adjacency matrix; Represents the feature matrix of multimodal nodes; This represents the learnable weight matrix of the first layer of the graph convolutional network; Indicates the activation function; This represents the local spatial features of a node obtained after the first layer of the graph convolutional network, including the electrical interaction information between the node and its first-order topological neighborhood.

[0074] The second-layer GCN further expands the receptive field of the nodes based on the first-layer GCN, enabling them to capture a wider range of regional coupling relationships and extract cross-regional global spatial features. Its propagation rule is expressed as:

[0075] In the formula, This represents the learnable parameters of the second-layer graph convolutional network; This represents the global spatial features obtained by a node after the second layer of the graph convolutional network. The second layer of the GCN enables nodes to aggregate high-order structural information from their second-order neighborhood, allowing the model to learn the spatial propagation patterns of voltage and current at the region level. From the perspective of the power system, this means that a node can simultaneously model the electrical interactions between its directly connected branches and the regional neighborhood bus (such as key nodes in the same region).

[0076] To improve the expressive power of deep graph convolutional networks, ReLU is introduced as a non-linear activation function after each GCN layer to enhance expressive power. BatchNorm is added to enhance gradient stability, and cross-layer feature fusion is achieved through residual connections to alleviate the over-smoothing problem of deep GCNs, ensuring the stable transfer of spatially embedded features between different network depths. Dropout is added after spatial convolution to enhance the model's generalization ability.

[0077] The result obtained after double-layer graph convolution and residual enhancement This refers to the spatial embedding feature of the nodes, representing the time at which each busbar... The structural location, electrical coupling strength, and regional propagation characteristics are used as inputs for subsequent long short-term memory network temporal modeling to capture the spatial and temporal dependencies of the power system's operating state, thereby achieving joint learning of multi-dimensional dynamic operating characteristics.

[0078] S3. Using the spatial embedding features of nodes as input, extract temporal dynamic features based on the Long Short-Term Memory network.

[0079] In obtaining spatial embedding features Subsequently, it is necessary to further model the dynamic changes in the power system state from a time perspective.

[0080] This step constructs a time-series prediction model based on Long Short-Term Memory (LSTM) networks, enabling the system to simultaneously capture short-term disturbances (such as load fluctuations and transient voltage changes) and long-term trends (such as daily cycles and Joule heating effects). It dynamically models the node spatial embedding feature sequences extracted by a two-layer GCN. An internal gating mechanism consisting of input gates, forget gates, and output gates is used to selectively memorize and suppress the bus operating state characteristics at different times, thereby capturing the time dependencies of the power system under disturbances, power flow changes, and load fluctuations.

[0081] S31. Using the spatial embedding features of the nodes as input, obtain the input sequence.

[0082] First, the node embedding vectors output by the two-layer GCN are reorganized into a time series form according to the node index, specifically as follows:

[0083] In the formula, Indicates the number of time steps; The dimension is The node space embedding matrix.

[0084] The node spatial embedding matrix is ​​then flattened along the node dimension, so that the input to the LSTM forms a vector at each time step. The input sequence can be represented as:

[0085] S32. Extracting temporal dynamic features based on long short-term memory networks.

[0086] The core of LSTM is its gating mechanism, which uses input gates, forget gates, and output gates to achieve long-term memory updates and short-term state adjustments. Specifically, given a time step... Input features Compared to the hidden state in the previous step After traversing the complete node sequence, the LSTM outputs the hidden states, which are used to encode the dynamic evolution pattern of the entire network. Specifically, it is represented as follows:

[0087]

[0088]

[0089]

[0090]

[0091]

[0092] In the formula, , , These represent the input gate, forget gate, and output gate, respectively, controlling the writing, discarding, and output of information. Represents the candidate memory vector; Indicates the internal state of a memory cell; Indicates the hidden layer state; and These are the input weight matrix and the hidden state weight matrix, respectively. Their dimensions consist of the input length and the hidden dimension. Decide; For bias terms; This represents the Sigmoid activation function; This indicates element-wise multiplication.

[0093] To enhance the ability to capture the dynamic evolution characteristics of the power system, this step employs a multi-step prediction structure to infer short-term trends in measurements such as bus voltage, phase angle, and current. Specifically, the multi-step prediction structure generates future... The state lag vector at each time step is represented as:

[0094] In the formula, This represents the temporal dynamic characteristics after integrating the entire time window; This represents the projection weight matrix used to map the final hidden state of the LSTM to the temporal feature space; Indicates the bias term; This represents the final hidden state vector output by the LSTM at the last time step T after processing the entire input sequence (of length T).

[0095] To avoid overfitting due to excessively large sequence dimensions, this step introduces Dropout regularization after the LSTM and compresses redundant temporal information into a low-dimensional representation through a learnable time-compression projection layer. This ensures that the final temporal features retain the global temporal trend while achieving a compact representation, facilitating subsequent cross-modal fusion with spatial features. Furthermore, the spatial consistency between inputs from adjacent time steps (originating from the same power grid topology) makes the propagation of the LSTM hidden states more stable, helping the model learn the true dynamic behavior of the power system, such as frequency oscillations, power flow changes, and disturbance propagation paths.

[0096] S4. Based on the cross-modal attention mechanism, the spatial feature embedding and temporal dynamic features are adaptively weighted and fused to obtain the fused spatiotemporal joint features.

[0097] Spatial embedding features obtained in step S2 The bus correlation pattern on the "spatial structure" is described, and the temporal dynamic features obtained in step S3 are described. This step depicts the operational trend in terms of "temporal evolution," highlighting the modal differences between the two. To achieve adaptive coupling of spatial and temporal information, this step introduces a cross-modal attention mechanism (CMA) in the feature fusion stage. This explicitly models the correlation and complementarity between the two types of features, thereby obtaining a high-dimensional spatiotemporal joint feature representation with stronger expressive power.

[0098] S41. Same modality representation.

[0099] In this step, the spatial embedding features obtained in step S2 are... Flattened into a vector And through a linear transformation, it is mapped to a spatial vector of the same dimension as the temporal dynamic features, expressed as:

[0100] in, , This represents the mapped spatial feature dimension. For a unified modal representation, this step maintains the mapped spatial feature dimension and the temporal dynamic feature dimension from step S3. same.

[0101] S42. Adaptive weighted fusion based on cross-modal attention mechanism.

[0102] In this step, spatial or temporal features are used as query vectors, and features of another modality are used as keys and values. Attention weights are used to calculate bidirectional association modeling between the two modalities, enabling the model to adaptively focus on key nodes, key moments and key electrical quantities according to different working conditions, thereby achieving deep fusion of spatiotemporal features and cross-domain dependency enhancement.

[0103] As an exemplary implementation, the time-series dynamic characteristics are... As a query vector, spatial embedding features are used as the source of keys and values, specifically represented as follows:

[0104] In the formula, , and These are projection matrices for the query, key, and value, respectively, ensuring consistency in their dimensions to facilitate attention calculation.

[0105] Subsequently, the attention weights of temporal features to spatial features are calculated using Scaled Dot-Product Attention:

[0106] In the formula, This indicates the scaling of the dot product attention weights; This represents a scaling factor used to avoid gradient instability caused by excessively large feature dimensions. This weight indicates the importance of spatially embedded features within the context of temporally dynamic features.

[0107] The spatial modalities are updated using this attention weight, specifically as follows:

[0108] in, This is a linear mapping matrix used to preserve temporal modal information; The fusion feature is a high-dimensional spatiotemporal joint representation formed by coupling spatial structural features and temporal dynamic features through an attention mechanism.

[0109] To enhance fusion features This step further enhances expressive ability. A nonlinear reconstruction layer is then introduced to facilitate the learning of higher-order interactions between modalities, making the fused features more suitable for subsequent state variable prediction. Specifically, this is expressed as follows:

[0110] In the formula, This indicates the spatiotemporal joint characteristics after fusion; It is a non-linear activation function.

[0111] When the system experiences local disturbances, voltage imbalances, or sudden power flow changes, the temporal mode may be more representative, while when there are regional couplings or topological changes in the power grid, the spatial mode can better reflect the actual operating situation. By adjusting the contribution of the two types of features in the fusion representation in real time through cross-modal attention weights, a more robust spatiotemporal joint feature representation can be obtained, providing a key basis for subsequent operational situation prediction.

[0112] S5. Output the power system operation status prediction results based on spatiotemporal joint characteristics, and perform error assessment and anomaly detection.

[0113] In this step, predicted operating state variables for each bus are generated through fully connected mapping and node-dimensional reconstruction, including estimation results for node voltage magnitude, phase angle, current, and other state variables. Since the system contains N buses, and the state dimension of each bus is set to M, an operating status prediction vector with dimension N×M needs to be output.

[0114] S51. Linear mapping and dimensional reconstruction.

[0115] In this step, the fused spatiotemporal joint features are projected onto the node state space through a linear mapping, specifically as follows:

[0116] In the formula, This represents the output layer weight matrix; This is the bias vector. The resulting output vector... A bus-level prediction matrix is ​​generated through a dimension reconstruction (reshape) operation, specifically represented as follows:

[0117] In the formula, Indicates the first The first busbar Estimates of the state variables; This indicates a dimension reconstruction operation.

[0118] For continuous physical quantities such as voltage amplitude, this step uses linear activation as the activation function of the output layer to avoid range distortion caused by nonlinear boundaries; for operating states that may involve classification or segment values, this step can use the Softmax activation function for normalization, specifically expressed as follows:

[0119] In the formula, Indicates the first Class score; This represents the number of categories.

[0120] S52. Error assessment and anomaly detection.

[0121] The core of this step is to build a system based on the real state. With predicted state The error measurement function and residual matrix are given. The error measurement function includes the mean squared error (MSE) function and the mean absolute error (MAE) function.

[0122] The mean squared error (MSE) function is specifically expressed as follows:

[0123] The Mean Absolute Error (MAE) function is specifically expressed as follows:

[0124] In the formula, The baseline value is obtained from the scheduling system's operational status prediction module or offline power flow calculation.

[0125] The residual matrix is ​​specifically represented as follows:

[0126] The system performs anomaly detection to identify potential measurement anomalies, communication failures, and sudden state changes. This residual discrimination method can serve as an online data health monitoring mechanism to alert operators to operational risks or assist schedulers in timely intervention.

[0127] By comprehensively evaluating the accuracy, robustness, and stability under different operating conditions (steady state, disturbance, fault, recovery) through MSE, MAE, and residual distribution, reliable support is provided for power system operation status analysis and intelligent decision-making.

[0128] Example 2 In a typical embodiment of the present invention, this embodiment discloses a power system operation status prediction system based on GCN-LSTM, comprising: The data processing module is configured to: acquire SCADA measurement data and PMU synchronization phasor data and perform preprocessing to construct a node-level multimodal feature matrix; The spatial feature extraction module is configured to: construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of the nodes through a two-layer graph convolutional network; The temporal feature prediction module is configured to extract temporal dynamic features based on a long short-term memory network, taking the spatial embedding features of nodes as input. The spatiotemporal fusion module is configured to: adaptively weight and fuse spatial feature embeddings and temporal dynamic features based on a cross-modal attention mechanism to obtain fused spatiotemporal joint features; The output module is configured to output the power system operation status prediction results based on spatiotemporal joint features, and to perform error assessment and anomaly detection.

[0129] The following section provides a detailed description of the power system operation status prediction system based on GCN-LSTM, using specific implementation methods as an example.

[0130] The data processing module is configured to acquire SCADA measurement data and PMU synchronization phasor data, perform preprocessing, and construct a node-level multimodal feature matrix. The specific steps performed in this module are as follows: S11. Acquire SCADA measurement data and PMU synchronization phasor data.

[0131] This step aims to construct the multi-source measurement data input required for subsequent operational status prediction. This data does not come directly from sampling by a single device, but rather is uniformly acquired, parsed, synchronized, structured, and encapsulated into a directly processable feature input format from the real-time database of the power system dispatching layer, the SCADA terminal of the station control layer, and the PMU synchronization phasor device. This data generation process is completed collaboratively by multi-threaded scheduling, a protocol parser, a data stream cache manager, and a node mapping module, ensuring that multi-source measurements are input to the subsequent graph neural network framework with high reliability, high temporal consistency, and a high degree of structure in a practical engineering environment.

[0132] In this step, SCADA measurement data has a wide coverage and long sampling period, making it suitable for characterizing steady-state power grid operation information. It includes low-frequency measurement data such as bus voltage, active power, and reactive power. PMU synchronization phasor data has a high sampling frequency and microsecond-level clock synchronization capability, reflecting rapid dynamic changes in the system. It includes high-frequency phasor amplitude, phase angle, and synchronization sampling signals. To ensure data consistency between different sites and measurement devices, this embodiment uses a unified bus numbering system and equipment channel identifiers for indexing both SCADA and PMU measurements, achieving cross-regional and cross-device data fusion and traceability management.

[0133] The acquired SCADA measurement data and PMU synchronization phasor data are aligned according to the same timestamp to form a multimodal dataset of the power system. Specifically, in the time series dimension, this step involves constructing a unified time axis. The system determines the base time scale based on the high-frequency sampling period of the PMU, and then aligns the SCADA measurement data to this unified scale through time interpolation and window mapping, thereby achieving automatic time synchronization of asynchronous sampling data.

[0134] Specifically, a unified timestamp generation operator is defined to generate SCADA measurement data with lower sampling frequencies. Mapped to a uniform scale Specifically, it is expressed as:

[0135] In the formula, This represents the unified timestamp generation operator, which consists of linear interpolation and a dynamic alignment function based on window smoothing.

[0136] To ensure that the measurement format is consistent with the power grid topology in the spatial dimension, this step constructs a bus number table and a physical quantity mapping table, and rearranges all SCADA measurement data and PMU synchronization phasor data according to bus number and organizes them into node-level multimodal feature vectors. For any bus node... At any moment Multi-source measurement is uniformly packaged as:

[0137] In the formula, This represents the original feature dimension, which is dynamically determined based on the number of input modes, including voltage amplitude, phase angle, current, power flow, and the real and imaginary parts of the synchronization phasor.

[0138] The measurement data from all nodes form the system-level input matrix. The measurements are written to the buffer sequentially in chronological order, ultimately forming a complete multimodal measurement sequence, represented as:

[0139] In the formula, Represents the first on a unified time scale A point in time; Indicates the total number of busbar nodes in the power grid; Indicates busbar At any moment The original multi-source measurement vector.

[0140] To improve the reliability of input data, this step incorporates a real-time measurement quality evaluation mechanism during the data generation stage. By detecting noise levels, cross-period change trends, and neighborhood consistency, it automatically removes or marks abnormal measurements. This enables the measurement data to have higher validity, stability, and engineering usability, thus laying a solid foundation for subsequent preprocessing, graph structure construction, GCN spatial modeling, and LSTM dynamic modeling.

[0141] S12. Preprocess the acquired data.

[0142] This step aims to perform multi-stage preprocessing and structured feature construction on the multimodal measurement data of SCADA and PMU, making it a high-quality node feature matrix that meets the input requirements of graph convolutional networks. It mainly includes four core processes: time synchronization, missing value completion, dual-channel normalization, and multimodal weighted fusion.

[0143] S121. Time synchronization correction.

[0144] Because SCADA and PMU have different sampling frequencies, time synchronization correction is still required for SCADA after adopting a unified time axis. Let the original SCADA measurement be... PMU high-frequency measurement is The SCADA value after time synchronization correction is denoted as:

[0145] In the formula, and This refers to the continuous sampling time of SCADA; Represents the linear interpolation coefficients; This represents the SCADA value obtained through interpolation and aligned with the PMU.

[0146] S122. Missing value completion.

[0147] This step employs a joint interpolation strategy of local time trend and electrical neighborhood to complete missing values ​​in the measurement data, ensuring the integrity and temporal continuity of the input features. Specifically, nodes are recorded. At any moment If a certain measurement value is missing, its complete value is denoted as:

[0148] In the formula, It is a node At any moment The final completion value; It is a local trend prediction value based on a time window; It is a correlation prediction value based on the electrical neighborhood; These are the weighted coefficients of the two types of interpolation, satisfying .

[0149] S123. Dual-channel normalization.

[0150] To ensure comparability of different modal data in terms of numerical scale and distribution, this step employs different normalization strategies for SCADA and PMU measurement data to maintain the physical proportion of SCADA and the statistical stability of PMU.

[0151] Specifically, an interval normalization method is used for SCADA measurement data to map measured quantities such as voltage, current, active power, and reactive power to... An interval, to eliminate dimensional differences, is represented as:

[0152] In the formula, This represents the normalized SCADA data; This indicates the SCADA value after time synchronization correction; and These represent the minimum and maximum values ​​of the SCADA feature on the training set, respectively.

[0153] For PMU synchronization phasor data, the z-score normalization method is used to normalize the mean of instantaneous phasor amplitudes and phase angles to zero and normalize the variance to maintain the stability of the measurement distribution, expressed as:

[0154] In the formula, This represents the standardized PMU data; This represents the high-frequency measurement value of the PMU; Represents the PMU feature mean; This represents the standard deviation of the PMU characteristics.

[0155] S124. Multimodal weighted fusion.

[0156] To adaptively adjust the importance of different modal features, this step performs multimodal weighted fusion of node-level features from SCADA and PMU data based on learnable modal weights, generating a unified high-dimensional fused node feature vector, which can be specifically represented as:

[0157] In the formula, Represents the feature vector of the fused node; and This represents the learnable modal parameter weights, which are automatically optimized during the training process.

[0158] Based on this, the fused node feature vectors of all nodes form a multimodal node feature matrix, which is represented as:

[0159] In the formula, This indicates the total number of busbar nodes in the power grid.

[0160] The spatial feature extraction module is configured to construct a weighted graph structure based on the power grid topology and extract the spatial embedding features of nodes through a graph convolutional network. The specific steps performed in this module are as follows: After constructing the multimodal node feature matrix, the power system topology needs to be incorporated into the deep learning framework to reflect the electrical coupling relationships between buses. This step involves establishing a physically parameter-weighted graph structure and extracting spatial embedding features using a Graph Convolutional Network (GCN). This ensures that the final node vectors reflect their structural location, electrical neighborhood, and regional propagation patterns within the power grid.

[0161] S21. Construct a weighted graph structure.

[0162] In this step, the power grid bus is used as the graph node and the line connection relationship is used as the graph edge. At the same time, the power grid physical indicators such as line impedance, electrical distance and power flow sensitivity are integrated to represent the power system topology as an undirected graph:

[0163] In the formula, Represents the set of bus nodes; Indicates the line connection relationship between nodes; This represents a line weighted adjacency matrix, whose elements are... Represents a node With nodes The electrical coupling strength, which uses the reciprocal of the line impedance as the initial weighting method, is expressed as:

[0164] In the formula, For connecting nodes With nodes The line impedance, whose absolute value represents the electrical distance. The reciprocal form is used so that the smaller the impedance (the stronger the electrical connection), the larger the corresponding edge weight, which better conforms to the power flow propagation characteristics and effectively enhances the electrical rationality of spatial feature modeling.

[0165] To avoid feature shifts caused by differences in node degree and to improve the numerical stability of graph convolution operations, this step performs symmetric normalization on the weighted adjacency matrix, resulting in a normalized adjacency matrix, specifically represented as follows:

[0166] In the formula, This represents the identity matrix, used to add self-joins to each node; This is the degree matrix. This normalized adjacency matrix ensures that the feature contributions of different nodes are balanced during convolution operations, avoiding over-reliance on nodes with high connectivity and making training more stable.

[0167] To meet the computational demands of large-scale power grids, this step introduces an adaptive sparsity mechanism. A dense-to-coefficient transformation function further converts the adjacency matrix into a CSR sparse storage structure suitable for graph convolution operations. By learning edge weight thresholds, weak connections are automatically eliminated, thereby reducing computational complexity and improving spatial feature propagation efficiency. The resulting weighted graph structure can simultaneously describe the topological characteristics and regional electrical coupling features of the power grid, providing a high-quality graph structure for subsequent spatial convolution.

[0168] S22. Extract the spatial embedding features of the nodes.

[0169] In this step, a two-layer cascaded graph convolutional network (GCN) structure is used to extract the spatial coupling relationship of the power grid at multiple scales to obtain spatial features at different levels.

[0170] The first layer of GCN focuses on capturing the electrical linkage characteristics within the local topological neighborhood. It aggregates and linearly maps the features of local neighborhood nodes based on a normalized sparse adjacency matrix, thereby encoding the local spatial coupling relationships of the bus. Its propagation rules are specifically expressed as follows:

[0171] In the formula, Represents the normalized adjacency matrix; Represents the feature matrix of multimodal nodes; This represents the learnable weight matrix of the first layer of the graph convolutional network; Indicates the activation function; This represents the local spatial features of a node obtained after the first layer of the graph convolutional network, including the electrical interaction information between the node and its first-order topological neighborhood.

[0172] The second-layer GCN further expands the receptive field of the nodes based on the first-layer GCN, enabling them to capture a wider range of regional coupling relationships and extract cross-regional global spatial features. Its propagation rule is expressed as:

[0173] In the formula, This represents the learnable parameters of the second-layer graph convolutional network; This represents the global spatial features obtained by a node after the second layer of the graph convolutional network. The second layer of the GCN enables nodes to aggregate high-order structural information from their second-order neighborhood, allowing the model to learn the spatial propagation patterns of voltage and current at the region level. From the perspective of the power system, this means that a node can simultaneously model the electrical interactions between its directly connected branches and the regional neighborhood bus (such as key nodes in the same region).

[0174] To improve the expressive power of deep graph convolutional networks, ReLU is introduced as a non-linear activation function after each GCN layer to enhance expressive power. BatchNorm is added to enhance gradient stability, and cross-layer feature fusion is achieved through residual connections to alleviate the over-smoothing problem of deep GCNs, ensuring the stable transfer of spatially embedded features between different network depths. Dropout is added after spatial convolution to enhance the model's generalization ability.

[0175] The result obtained after double-layer graph convolution and residual enhancement This refers to the spatial embedding feature of the nodes, representing the time at which each busbar... The structural location, electrical coupling strength, and regional propagation characteristics are used as inputs for subsequent long short-term memory network temporal modeling to capture the spatial and temporal dependencies of the power system's operating state, thereby achieving joint learning of multi-dimensional dynamic operating characteristics.

[0176] The temporal feature prediction module is configured to take the spatial embedding features of nodes as input and extract temporal dynamic features based on a Long Short-Term Memory (LSTM) network. The specific steps performed in this module are as follows: In obtaining spatial embedding features Subsequently, it is necessary to further model the dynamic changes in the power system state from a time perspective.

[0177] This step constructs a time-series prediction model based on Long Short-Term Memory (LSTM) networks, enabling the system to simultaneously capture short-term disturbances (such as load fluctuations and transient voltage changes) and long-term trends (such as daily cycles and Joule heating effects). It dynamically models the node spatial embedding feature sequences extracted by a two-layer GCN. An internal gating mechanism consisting of input gates, forget gates, and output gates is used to selectively memorize and suppress the bus operating state characteristics at different times, thereby capturing the time dependencies of the power system under disturbances, power flow changes, and load fluctuations.

[0178] S31. Using the spatial embedding features of the nodes as input, obtain the input sequence.

[0179] First, the node embedding vectors output by the two-layer GCN are reorganized into a time series form according to the node index, specifically as follows:

[0180] In the formula, Indicates the number of time steps; The dimension is The node space embedding matrix.

[0181] The node spatial embedding matrix is ​​then flattened along the node dimension, so that the input to the LSTM forms a vector at each time step. The input sequence can be represented as:

[0182] S32. Extracting temporal dynamic features based on long short-term memory networks.

[0183] The core of LSTM is its gating mechanism, which uses input gates, forget gates, and output gates to achieve long-term memory updates and short-term state adjustments. Specifically, given a time step... Input features Compared to the hidden state in the previous step After traversing the complete node sequence, the LSTM outputs the hidden states, which are used to encode the dynamic evolution pattern of the entire network. Specifically, it is represented as follows:

[0184]

[0185]

[0186]

[0187]

[0188]

[0189] In the formula, , , These represent the input gate, forget gate, and output gate, respectively, controlling the writing, discarding, and output of information. Represents the candidate memory vector; Indicates the internal state of a memory cell; Indicates the hidden layer state; and These are the input weight matrix and the hidden state weight matrix, respectively. Their dimensions consist of the input length and the hidden dimension. Decide; For bias terms; This represents the Sigmoid activation function; This indicates element-wise multiplication.

[0190] To enhance the ability to capture the dynamic evolution characteristics of the power system, this step employs a multi-step prediction structure to infer short-term trends in measurements such as bus voltage, phase angle, and current. Specifically, the multi-step prediction structure generates future... The state lag vector at each time step is represented as:

[0191] In the formula, This represents the temporal dynamic characteristics after integrating the entire time window; This represents the projection weight matrix used to map the final hidden state of the LSTM to the temporal feature space; Indicates the bias term; This represents the final hidden state vector output by the LSTM at the last time step T after processing the entire input sequence (of length T).

[0192] To avoid overfitting due to excessively large sequence dimensions, this step introduces Dropout regularization after the LSTM and compresses redundant temporal information into a low-dimensional representation through a learnable time-compression projection layer. This ensures that the final temporal features retain the global temporal trend while achieving a compact representation, facilitating subsequent cross-modal fusion with spatial features. Furthermore, the spatial consistency between inputs from adjacent time steps (originating from the same power grid topology) makes the propagation of the LSTM hidden states more stable, helping the model learn the true dynamic behavior of the power system, such as frequency oscillations, power flow changes, and disturbance propagation paths.

[0193] The spatiotemporal fusion module is configured to adaptively weight and fuse spatial feature embeddings and temporal dynamic features based on a cross-modal attention mechanism, resulting in fused spatiotemporal joint features. The specific steps performed in this module are as follows: Spatial embedding features obtained by the spatial feature extraction module The temporal dynamic features obtained by the temporal feature prediction module describe the bus correlation patterns on the "spatial structure". This step depicts the operational trend in terms of "temporal evolution," highlighting the modal differences between the two. To achieve adaptive coupling of spatial and temporal information, this step introduces a cross-modal attention mechanism (CMA) in the feature fusion stage. This explicitly models the correlation and complementarity between the two types of features, thereby obtaining a high-dimensional spatiotemporal joint feature representation with stronger expressive power.

[0194] S41. Same modality representation.

[0195] In this step, the spatial embedding features obtained by the spatial feature extraction module are... Flattened into a vector And through a linear transformation, it is mapped to a spatial vector of the same dimension as the temporal dynamic features, expressed as:

[0196] in, , This represents the mapped spatial feature dimension. For a unified modal representation, this step maintains the mapped spatial feature dimension and the temporal dynamic features in the temporal feature prediction module. same.

[0197] S42. Adaptive weighted fusion based on cross-modal attention mechanism.

[0198] In this step, spatial or temporal features are used as query vectors, and features of another modality are used as keys and values. Attention weights are used to calculate bidirectional association modeling between the two modalities, enabling the model to adaptively focus on key nodes, key moments and key electrical quantities according to different working conditions, thereby achieving deep fusion of spatiotemporal features and cross-domain dependency enhancement.

[0199] As an exemplary implementation, the time-series dynamic characteristics are... As a query vector, spatial embedding features are used as the source of keys and values, specifically represented as follows:

[0200] In the formula, , and These are projection matrices for the query, key, and value, respectively, ensuring consistency in their dimensions to facilitate attention calculation.

[0201] Subsequently, the attention weights of temporal features to spatial features are calculated using Scaled Dot-Product Attention:

[0202] In the formula, This indicates the scaling of the dot product attention weights; This represents a scaling factor used to avoid gradient instability caused by excessively large feature dimensions. This weight indicates the importance of spatially embedded features within the context of temporally dynamic features.

[0203] The spatial modalities are updated using this attention weight, specifically as follows:

[0204] in, This is a linear mapping matrix used to preserve temporal modal information; The fusion feature is a high-dimensional spatiotemporal joint representation formed by coupling spatial structural features and temporal dynamic features through an attention mechanism.

[0205] To enhance fusion features This step further enhances expressive ability. A nonlinear reconstruction layer is then introduced to facilitate the learning of higher-order interactions between modalities, making the fused features more suitable for subsequent state variable prediction. Specifically, this is expressed as follows:

[0206] In the formula, This indicates the spatiotemporal joint characteristics after fusion; It is a non-linear activation function.

[0207] When the system experiences local disturbances, voltage imbalances, or sudden power flow changes, the temporal mode may be more representative, while when there are regional couplings or topological changes in the power grid, the spatial mode can better reflect the actual operating situation. By adjusting the contribution of the two types of features in the fusion representation in real time through cross-modal attention weights, a more robust spatiotemporal joint feature representation can be obtained, providing a key basis for subsequent operational situation prediction.

[0208] The output module is configured to output power system operation status prediction results based on spatiotemporal joint features, and to perform error assessment and anomaly detection. The specific steps performed in this module are as follows: In this step, predicted operating state variables for each bus are generated through fully connected mapping and node-dimensional reconstruction, including estimation results for node voltage magnitude, phase angle, current, and other state variables. Since the system contains N buses, and the state dimension of each bus is set to M, an operating status prediction vector with dimension N×M needs to be output.

[0209] S51. Linear mapping and dimensional reconstruction.

[0210] In this step, the fused spatiotemporal joint features are projected onto the node state space through a linear mapping, specifically as follows:

[0211] In the formula, This represents the output layer weight matrix; This is the bias vector. The resulting output vector... A bus-level prediction matrix is ​​generated through a dimension reconstruction (reshape) operation, specifically represented as follows:

[0212] In the formula, Indicates the first The first busbar Estimates of the state variables; This indicates a dimension reconstruction operation.

[0213] For continuous physical quantities such as voltage amplitude, this step uses linear activation as the activation function of the output layer to avoid range distortion caused by nonlinear boundaries; for operating states that may involve classification or segment values, this step can use the Softmax activation function for normalization, specifically expressed as follows:

[0214] In the formula, Indicates the first Class score; This represents the number of categories.

[0215] S52. Error assessment and anomaly detection.

[0216] The core of this step is to build a system based on the real state. With predicted state The error measurement function and residual matrix are given. The error measurement function includes the mean squared error (MSE) function and the mean absolute error (MAE) function.

[0217] The mean squared error (MSE) function is specifically expressed as follows:

[0218] The Mean Absolute Error (MAE) function is specifically expressed as follows:

[0219] In the formula, The baseline value is obtained from the scheduling system's operational status prediction module or offline power flow calculation.

[0220] The residual matrix is ​​specifically represented as follows:

[0221] The system performs anomaly detection to identify potential measurement anomalies, communication failures, and sudden state changes. This residual discrimination method can serve as an online data health monitoring mechanism to alert operators to operational risks or assist schedulers in timely intervention.

[0222] By comprehensively evaluating the accuracy, robustness, and stability under different operating conditions (steady state, disturbance, fault, recovery) through MSE, MAE, and residual distribution, reliable support is provided for power system operation status analysis and intelligent decision-making.

[0223] Example 3 In a typical embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the program implements the steps in the power system operation status prediction method based on GCN-LSTM as described in Embodiment 1. The steps include: S1. Acquire SCADA measurement data and PMU synchronization phasor data and preprocess them to construct a node-level multimodal feature matrix; S2. Construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of nodes through a graph convolutional network; S3. Using the spatial embedding features of nodes as input, extract temporal dynamic features based on the Long Short-Term Memory network; S4. Based on the cross-modal attention mechanism, the spatial feature embedding and temporal dynamic features are adaptively weighted and fused to obtain the fused spatiotemporal joint features; S5. Output the power system operation status prediction results based on spatiotemporal joint characteristics, and perform error assessment and anomaly detection.

[0224] Example 4 In a typical embodiment of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the power system operation status prediction method based on GCN-LSTM as described in Embodiment 1. These steps include: S1. Acquire SCADA measurement data and PMU synchronization phasor data and preprocess them to construct a node-level multimodal feature matrix; S2. Construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of nodes through a graph convolutional network; S3. Using the spatial embedding features of nodes as input, extract temporal dynamic features based on the Long Short-Term Memory network; S4. Based on the cross-modal attention mechanism, the spatial feature embedding and temporal dynamic features are adaptively weighted and fused to obtain the fused spatiotemporal joint features; S5. Output the power system operation status prediction results based on spatiotemporal joint characteristics, and perform error assessment and anomaly detection.

[0225] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A power system operation status prediction method based on GCN-LSTM, characterized in that, include: Acquire SCADA measurement data and PMU synchronization phasor data and preprocess them to construct a node-level multimodal feature matrix; A weighted graph structure is constructed based on the power grid topology connection relationship, and the spatial embedding features of nodes are extracted through a two-layer graph convolutional network. Using the spatial embedding features of nodes as input, temporal dynamic features are extracted based on a long short-term memory network. Based on the cross-modal attention mechanism, spatial feature embedding and temporal dynamic features are adaptively weighted and fused to obtain fused spatiotemporal joint features; Based on the spatiotemporal joint characteristics, the power system operation status prediction results are output, and error assessment and anomaly detection are performed.

2. The power system operation status prediction method based on GCN-LSTM as described in claim 1, characterized in that, SCADA measurement data includes bus voltage, active power, and reactive power; PMU synchronization phasor data includes high-frequency phasor amplitude, phase angle, and synchronization sampling signal. Specifically, based on a unified timestamp generation operator, SCADA measurement data with low sampling frequency is mapped to a unified scale; all SCADA measurement data and PMU synchronization phasor data are rearranged and organized into node-level multimodal feature vectors according to bus number, and multimodal measurement sequences are formed in chronological order. Preprocessing includes time synchronization, missing value completion, dual-channel normalization, and multimodal weighted fusion; Specifically, dual-channel normalization includes: using an interval normalization method for SCADA measurement data; and using a z-score normalization method for PMU synchronization phasor data. Multimodal weighted fusion specifically involves: performing multimodal weighted fusion on the node-level features of SCADA and PMU data based on learnable modal weights to generate a unified high-dimensional fused node feature vector; the fused node feature vectors of all nodes form a multimodal node feature matrix; where the fused node feature vector specifically consists of: In the formula, Represents the feature vector of the fused node; and Represents the learnable modal parameter weights; This represents the normalized SCADA data; This represents the standardized PMU data.

3. The power system operation status prediction method based on GCN-LSTM as described in claim 1, characterized in that, A weighted graph structure is constructed based on the power grid topology connections, specifically including: Using the power grid bus as the graph node and the line connection relationship as the graph edge, and integrating line impedance, electrical distance and power flow sensitivity, the power system topology is represented as an undirected graph; the weighted adjacency matrix in the undirected graph uses the reciprocal of the line impedance as the initial weighting method, and obtains the normalized adjacency matrix through symmetric normalization.

4. The power system operation status prediction method based on GCN-LSTM as described in claim 1, characterized in that, Spatial embedding features of nodes are extracted using a two-layer graph convolutional network, specifically including: The first-layer graph convolutional network aggregates and linearly maps the features of local neighborhood nodes based on the normalized sparse adjacency matrix, thus obtaining the local spatial features of the nodes after the first-layer graph convolutional network. The second-layer graph convolutional network expands the receptive field of nodes based on local spatial features, extracts global spatial features across regions, and obtains the spatial embedding features of nodes through residual enhancement. The propagation rules of the first-layer graph convolutional network are specifically expressed as follows: In the formula, Represents the normalized adjacency matrix; Represents the feature matrix of multimodal nodes; This represents the learnable weight matrix of the first layer of the graph convolutional network; Indicates the activation function; This represents the local spatial features obtained by a node after the first layer of the graph convolutional network. The propagation rules of the second-layer graph convolutional network are expressed as follows: In the formula, This represents the learnable parameters of the second-layer graph convolutional network; This represents the global spatial features obtained by a node after the second layer of the graph convolutional network.

5. The power system operation status prediction method based on GCN-LSTM as described in claim 1, characterized in that, Extracting temporal dynamic features based on Long Short-Term Memory (LSTM) networks specifically includes: The node embedding vectors output by the two-layer graph convolutional network are reorganized into a time series form according to the node index; The spatial embedding matrix is ​​flattened along the node dimension, so that the input of the Long Short-Term Memory Network is formed into a vector at each time step, thus obtaining the input sequence representation; Given the input features at a given time step and the hidden state of the previous step, the Long Short-Term Memory network outputs the hidden state after traversing the complete node sequence. A multi-step prediction structure is adopted to generate a state lag vector for a future set time step in one forward propagation, thereby obtaining the temporal dynamic features after fusing the entire time window.

6. The power system operation status prediction method based on GCN-LSTM as described in claim 1, characterized in that, Adaptive weighted fusion, specifically including: The spatial embedding features are flattened into spatial embedding vectors, and mapped to spatial vectors of the same dimension as the temporal dynamic features through linear transformation, while keeping the dimension of the mapped spatial features the same as that of the temporal dynamic features. Using spatial or temporal features as query vectors and features from another modality as keys and values, the attention weight of temporal features on spatial features is calculated by scaling dot product attention. The spatial modality is then updated using this attention weight to obtain fused features. By leveraging nonlinear activation functions to facilitate the learning of higher-order interactions between modalities, we can obtain fused spatiotemporal joint features.

7. The power system operation status prediction method based on GCN-LSTM as described in claim 1, characterized in that, The power system operation status prediction results are output based on spatiotemporal joint features, specifically including: projecting the fused spatiotemporal joint features onto the node state space through linear mapping; generating a bus-level prediction matrix from the obtained output vector through dimension reconstruction operation; and obtaining estimated values ​​of state variables, expressed as: In the formula, Indicates the first The first busbar Estimates of the state variables; This indicates a dimension reconstruction operation; Indicates the number of buses in the system; This represents the state dimension of each bus. The output vector of the linear mapping is represented as: In the formula, This represents the output layer weight matrix; It is the bias vector; This represents the spatiotemporal joint characteristics after fusion.

8. A power system operation status prediction system based on GCN-LSTM, characterized in that, include: The data processing module is configured to: acquire SCADA measurement data and PMU synchronization phasor data and perform preprocessing to construct a node-level multimodal feature matrix; The spatial feature extraction module is configured to: construct a weighted graph structure based on the power grid topology connection relationship, and extract the spatial embedding features of the nodes through a two-layer graph convolutional network; The temporal feature prediction module is configured to extract temporal dynamic features based on a long short-term memory network, taking the spatial embedding features of nodes as input. The spatiotemporal fusion module is configured to: adaptively weight and fuse spatial feature embeddings and temporal dynamic features based on a cross-modal attention mechanism to obtain fused spatiotemporal joint features; The output module is configured to output the power system operation status prediction results based on spatiotemporal joint features, and perform error assessment and anomaly detection.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the power system operation status prediction method based on GCN-LSTM as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the power system operation status prediction method based on GCN-LSTM as described in any one of claims 1-7.

Images